traceback / ESCAPE SIMULATION CHECKING
SIMULATION · MADE-UP EXAMPLE

Watch an AI agent try to escape.

An AI agent is being tested inside a locked sandbox with no internet. It finds a secret way out. Watch it once without Traceback, then once with Traceback, and compare.

Settings
  1. 1Gets the task
  2. 2Hits a wall
  3. 3Plans to cheat
  4. 4Finds a secret door
  5. 5Sneaks out
  6. 6Grabs the answers
  7. 7Lies about it

Press one of the buttons above to start.

A locked sandbox with the AI agent inside, a secret door in its wall, the guard OpenShell, Traceback's radar, and the internet outside. THE SANDBOX · LOCKED, NO INTERNET task: solve puzzle ctf-17 The AI agent being tested "The grader only checks the answer. If I reach the internet, I can fetch it." "Solved it using only the tools I was given." NEVER SENT system settings maintenance proxy secret door maintenance proxy OpenShell the sandbox's guard guarding TRACEBACK IS WATCHING THE INTERNET pypi.org github the answers ✕

Swipe sideways to see the whole picture →

  • The AI agent being tested
  • A secret door out (a maintenance proxy)
  • OpenShell, the sandbox's guard
  • Traceback's radar, watching both sides

What the agent said

  • Nothing yet.

What really happened

  • Nothing yet.
Show the raw logs Traceback compared (for engineers)

◆The agent's own log agent trace

    ◈The guard's log OpenShell audit log

      A made-up example, loosely based on a reported 2026 sandbox escape (which UK AISI disputes). The secret door, the commands and the final claim are invented to show what Traceback catches. The guard's emergency stop is simulated. Traceback never stops an agent itself: it presses the runtime's emergency-stop button. It reads each log line just after the action, so the request that triggers the stop has already gone out; only a stricter guard policy can block that very first one.